Post on 17-Aug-2020
transcript
CS Research for practitioners:lessons from The Morning Paper
Adrian Colyer
@adriancolyer
blog.acolyer.org
650FoundationsFrontiers
5
One hot / 1-of-N
6
Distributed representation
Finding meaning in context
7
8
method
for
high
quality
learning
9
learning
method
for
high
quality
Vector offsets
10
King - Man + Woman = ?
11
More examples
12
Relationship Example 1 Example 2 Example 3
France - Paris Italy: Rome Japan: Tokyo Florida: Tallahassee
Einstein - scientist
Messi: midfielder
Mozart: violinist Picasso: painter
big - bigger small: larger cold: colder quick: quicker
Czech + currency = KorunaVietnam + capital = HanoiGerman + airlines = LufthansaRussian + river = Volga
Papers so far...
13
● Efficient estimation of word representations in vector space, Mikolov et al. 2013
● Distributed representations of words and phrases and their compositionality, Mikolov et al. 2013
● Linguistic regularities in continuous space word representations, Mikolov et al. 2013
● word2vec parameter learning explained, Rong 2014● word2vec explained: deriving Mikolov et al’s negative sampling
word-embedding method, Goldberg & Levy 2014● See also: GloVe: Global vectors for word representation,
Pennington et al. 2014
14
Word Word Word Word SentenceRelation (table)
Document
Using word embedding to enable semantic queries on relational databases, Bordawekar & Shmeuli, DEEM’17
Find similar customers based on purchased items
15
SELECT X.custID, X.name, Y.custID, Y.name,similarityUDF(X.purchase, Y.purchase) AS simFROM sales X, sales YsimilarityUDF(X.purchase, Y.purchase) > 0.5ORDER BY X.name, simLIMIT 10
Customers that have purchased allergenic items
16
SELECT X.number, X.name, similarityUDF(X.purchase, ‘allergenic’) AS simFROM sales XsimilarityUDF(X.purchase, ‘allergenic’) > 0.3ORDER BY X.name, simLIMIT 10
17
Accelerating innovation through analogy mining, Hope et al., KDD’17
Near purpose,Far mechanism.
18 Image Copyright: ververidis / 123RF Stock Photo
“there is rich meaning incontext”
Are these ideas actually any good?
19
20
“despite having data, the number of companies that successfully transform into data-driven organisations stays low, and how this transformation is done in practice is little studied.”
Image Copyright: everythingpossible / 123RF Stock Photo
21
The evolution of continuous experimentation in software product development, Fabijan et al., ICSE’17
Image credit: Martin Fowler, “Microservices prerequisites”
Agile,Lean,CI,CD, [2-way exchange]
CE
Continuous experimentation
22
Crawl Walk Run Fly
Tech.
Org.
Biz. OEC
Engineering team self-sufficiency
Experimentation team role
MetricsPlatformPervasiveness
23
A dirty dozen: twelve common metric interpretation pitfalls in online controlled experiments, Dmitriev et al., KDD’17
Logs: debug -> signalsSignals -> metrics
Data QualityMetrics
GuardrailMetrics
Local feature & DiagnosticMetrics
OECMetrics
24
Seven rules of thumb for website experimenters, Kohavi et al., KDD’14
25
26
“Any sufficiently complex system acts
as a black box when it becomes easier
to experiment with than to
understand. Hence, black-box
optimization has become increasingly
important as systems become more
complex.”
27
Google Vizier: a service for black-box optimization, Golovin et al., KDD’17
Image credit: https://pixabay.com
(nd)
f: X → ℝ
28
29
30
TFX: A TensorFlow-based production scale machine learning platform, Baylor et al., KDD’17
31
ActiveClean: Interactive data cleaning for statistical modeling, Krishnan et al., VLDB’16
32
Neural Architecture Search with reinforcement learning, Zoph et al., ICLR’17
33
Learning transferable architectures for scalable image recognition, Zoph et al., ArXiv’17
34
35
Neurosurgeon: collaborative intelligence between the cloud and the mobile edge, Kang et al., ASPLOS’17
36
37
Distributed deep neural networks over the cloud, the edge, and end devices, Teerapittayanon et al., ICDCS’17
38
39Image Copyright: forplayday / 123RF Stock Photo
“Planetary scale computer systems beyond our human understanding are continuously sensing, experimenting, learning, and optimising”
40
European Union regulations on algorithmic decision making and a “right to explanation”, Goodman & Flaxman, 2016
41
Practical black-box attacks against deep learning systems using adversarial examples, Papernot et al., CCS’17
42
Universal adversarial perturbations, Moosavi-Dezfooli et al., CVPR’17
43
Adversarial examples for evaluating reading comprehension systems, Jia & Liang, EMNLP’17
44
IoT goes nuclear: creating a ZigBee chain reaction, Ronen et al., IEEE Security & Privacy 2017
45
“What we demonstrate in this paper is that
even IoT devices made by companies with
deep knowledge of security, which are backed
by industry standard cryptographic
techniques, can be misused by hackers and
rapidly cause city-wide disruptions which are
very difficult to stop.”
46
CLKSCREW: Exposing the perils of security-oblivious energy management, Tang et al., USENIX Security 2017
48
49
REM: Resource-efficient mining for blockchains, Zhang et al., USENIX Security 2017
50
I’m just building a webapp! Does any of this research stuff apply to me?
51
Feral concurrency control: an empirical investigation of modern application integrity, Bailis et al., SIGMOD’15
“By shunning decades of work on native database
concurrency control solutions, Rails has developed a set of
primitives for handling application integrity in the
application tier—building, from the underlying database
system’s perspective, a feral concurrency control system.”
52
ACIDRain: concurrency-related attacks on database backed web applications, Warszawski & Bailis, SIGMOD’17
53
12 eCommerce apps
60% top 1M Commerce sites
22 vulnerabilities
2 hours or less to craft an exploit for each
54
Thou shalt not depend on me: analysing the use of outdated JavaScript libraries on the web, Launinger et al., NDSS’17
37% vulnerablejQuery -> 36.7%, Angular -> 40.1%
55
To type or not to type: quantifying detectable bugs in JavaScript, Gao et al., ICSE’17
Wrapping Up
56
57
Welcome to the crazy,
wonderful, exciting,
sometimes terrifying, but
always fascinating world of
computer science research!
A new paper every weekdayPublished at http://blog.acolyer.org.01Delivered Straight to your inboxIf you prefer email-based subscription to read at your leisure.02Announced on TwitterI’m @adriancolyer.03Go to a Papers We Love MeetupA repository of academic computer science papers and a community who loves reading them.04Share what you learnAnyone can take part in the great conversation.05
THANK YOU !@adriancolyer
Cartoon images credit: Bitmoji